dataset = CityscapeDataset()
dataset.load_cityscape(args.data_dir, "val", args.num_frames, args.fix)
dataset.prepare()
print("Image Count: {}".format(len(dataset.image_ids)))
resnet = ResNet(config=config)
flownet = FlowNet(config=config)
maskrcnn = MaskRCNN(config=config)
warp = Warp(config=config)
decision = Decision(config=config)
model_path = args.restore_from
resnet.load_weights(model_path, by_name=True)
flownet.load_weights(model_path, by_name=True)
maskrcnn.load_weights(model_path, by_name=True)
decision.load_weights(args.decision_from, by_name=True)
AP50s = []
APs = []
seg_step = 0
flow_step = 0
score = 0
if args.method == 2:
target = -args.target
else:
target = args.target
if args.is_save and not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
for image_id in range(len(dataset.image_ids)):
# Load image and ground truth data
image, image_meta, gt_class_id, gt_bbox, gt_mask =\
def main():
args = get_arguments()
print(args)
config = CityscapeConfig()
config.FLOW = args.flowmodel
config.BACKBONE = args.resnetmodel
config.IMAGE_SHAPE = [1024, 1024, 6]
config.POST_NMS_ROIS_INFERENCE = 500
#config.display()
resnet = ResNet(config=config)
flownet = FlowNet(config=config)
maskrcnn = MaskRCNN(config=config)
warp = Warp(config=config)
decision = Decision(config=config)
model_path = args.restore_from
resnet.load_weights(model_path, by_name=True)
flownet.load_weights(model_path, by_name=True)
maskrcnn.load_weights(model_path, by_name=True)
decision.load_weights(args.decision_from, by_name=True)
seg_step = 0
flow_step = 0
target = args.target
list_file = open(args.data_list, 'r')
if args.is_save and not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
for step in range(args.num_steps):
f1 = list_file.readline().split('\n')[0]
f1 = os.path.join(args.data_dir, f1)
current = np.expand_dims(skimage.io.imread(f1), 0)
current, image_metas, window = mold_inputs(config, current)
image = current[0] + config.MEAN_PIXEL
if step == 0:
seg_step += 1
key_P2, key_P3, key_P4, key_P5, key_P6 = resnet.keras_model.predict(
current)
key = current
P2, P3, P4, P5, P6 = key_P2, key_P3, key_P4, key_P5, key_P6
else:
images = np.concatenate([current, key], 3)
flow, flow_feature = flownet.keras_model.predict(images)
score = decision.keras_model.predict(flow_feature)[0][0]
print("step: {:4d} predict score: {:.3f} target: {:.2f}".format(
step, score, target))
if score < target:
if args.dynamic and target < 50:
target -= 0.5
seg_step += 1
key_P2, key_P3, key_P4, key_P5, key_P6 = resnet.keras_model.predict(
current)
key = current
P2, P3, P4, P5, P6 = key_P2, key_P3, key_P4, key_P5, key_P6
else:
if args.dynamic and target < 95:
target += 0.25
flow_step += 1
P2, P3, P4, P5, P6 = warp.predict(
[key_P2, key_P3, key_P4, key_P5, key_P6, flow])
inputs = [image_metas, P2, P3, P4, P5, P6]
result = maskrcnn.detect_molded(inputs)
# Save
if args.is_save:
save_name = args.save_dir + 'mask' + str(step) + '.png'
colors = np.array(label_colors) / 255.0
pred_img = visualize.display_instances(image,
result['rois'],
result['masks'],
result['class_ids'],
class_names,
result['scores'],
colors=colors,
save_name=save_name)
print("segmentation steps:", seg_step, "flow steps:", flow_step)
